new data format and 2025 videos

README.md

@@ -30,9 +30,9 @@ git clone https://huggingface.co/datasets/hltcoe/wikivideo
 ```
 I would also tmux this because it might take a while. 
 
-
 #### Step 3: Untar the videos
-In the `data/` folder, you will see places for `data/audios` and `data/videos`. You need to untar the videos and audios into these folders. The audios file is `audios.tar.gz` and the videos is `videos.tar.gz`. 
+##### Videos from MultiVENT2.0 (WikiVideo 2024)
+These videos range from 2015-2024.
 ```bash
 # untar the videos 
 tar -xvzf videos.tar.gz -C data/videos
@@ -40,6 +40,15 @@ tar -xvzf videos.tar.gz -C data/videos
 tar -xvzf audios.tar.gz -C data/audios
 ```
 
+##### Videos from the year 2025 (WikiVideo25)
+These videos are for the year 2025. They are also the videos used in the MAGMaR shared task at ACL 2026. 
+```bash
+# untar the videos 
+tar -xvzf videos_2025.tar.gz -C data/videos
+# untar the audios
+tar -xvzf audios_2025.tar.gz -C data/audios
+```
+
 #### Finish
 Now you should be done. You will see a `annotations` folder in the huggingface repo, but this also exists in the `data/` folder already in the `data/wikivideo` directory. 
 
@@ -49,42 +58,41 @@ In the `data/wikivideo` directory, you will find the file `final_data.json` whic
 ```json
 {
     "Wikipedia Title": {
-        "article": "The article text",
-        "query_id": "test_pt2_query_XXX",
+        "claims": [["claim1", "claim2", ...], ...], 
         "original_article": ["sent1", "sent2", ...],
-        "audio_lists": [[true, false, true, ...], ...],
-        "video_lists": [[true, false, true, ...], ...],
-        "ocr_lists": [[true, false, true, ...], ...],
-        "neither_lists": [[true, false, true, ...], ...],
-        "video_ocr_lists": [[true, false, true, ...], ...],
-        "claims": [["claim1", "claim2", ...], ...],
-        "videos":{
-            "video_id": {
+        "claims_to_supporting_videos": {
+            "claim1": {
+                "supporting_videos": ["video_id1", "video_id2", ...],
+                "videos_modalities": {
+                    "video_id1": {
+                        "video": true,
+                        "audio": false,
+                        "ocr": true
+                    },
+                    ...
+                }
+            }
+        },
+        "article": "The article text",
+        "query_id": "query id",
+        "videos": {
+            "video_id1": {
                 "anon_scale_id": "XXX",
                 "language": "english",
                 "video_type": "Professional | Edited | Diet Raw | Raw",
-                "relevance": 3,
-            }
+                "relevance": 3
         }
     },
     ...
 }
 ```
 In this json, you see that the top level key is the Wikipeda Article Title. Each other key is defined as follows:
-- `article`: This is the human written article on the topic using the video data. 
+- `claims`: The claims from the original article
+- `original_article`: The original article split into sentences
+- `claims_to_supporting_videos`: A mapping from each claim to the videos that support it along with the modalities present in each video
+- `article`: This is the human written article on the topic using the video data.
 - `query_id`: This is the query id for the article from the MultiVENT 2.0 dataset. This will be helpful when doing RAG experiments.
-- `original_article`: This is the original Wikipedia article from MegaWika 2.0. It is sentence tokenized. Each sentence corresponsed to an index in the audio_lists, video_lists, ocr_lists, neither_lists, video_ocr_lists, and claims lists.
-- `claims`: This is a list of claims that are in the article. Each index corresponds to a sentence in the original article. Then each index in the sublist corresponds to a human written claim from the sentence. These specific claims correspond to the boolean elements in the audio_lists, video_lists, ocr_lists, neither_lists, and video_ocr_lists.
-- `audio_lists`: This is a list of lists of booleans. Each index corresponds to a sentence in the original article. Then each index in the sublist corresponds to a claim. If the boolean is TRUE then the claim is supported by the audio, otherwise it is not.
-- `video_lists`: This is a list of lists of booleans. Each index corresponds to a sentence in the original article. Then each index in the sublist corresponds to a claim. If the boolean is TRUE then the claim is supported by the video, otherwise it is not.
-- `ocr_lists`: This is a list of lists of booleans. Each index corresponds to a sentence in the original article. Then each index in the sublist corresponds to a claim. If the boolean is TRUE then the claim is supported by the ocr, otherwise it is not.
-- `neither_lists`: This is a list of lists of booleans. Each index corresponds to a sentence in the original article. Then each index in the sublist corresponds to a claim. If the boolean is TRUE then the claim is supported by any of the modalities. 
-- `video_ocr_lists`: This is a list of lists of booleans. Each index corresponds to a sentence in the original article. Then each index in the sublist corresponds to a claim. If the boolean is TRUE then the claim is supported by the video or ocr, otherwise it is not.
-- `videos`: This is a dictionary of videos relevant to the query topic. The key is the video id and the value is a dictionary of video metadata. The metadata is defined as follows:
-    - `anon_scale_id`: This is the anonnoynmous ID used in MultiVENT 2.0 for the video. This will help you deduplicate the test set of the dataset when doing video retrieval in the RAG experiments.  
-    - `language`: This is the language of the video. 
-    - `video_type`: This is the type of video. 
-    - `relevance`: This is the relevance of the video to the article. These values range from 0-3.
+- `videos`: Metadata about each video used in the article
 
 # RAG Data
 The videos that we used as the distractor set (and also include these videos in videos.tar.gz) can be found here MultiVENT 2.0 (https://huggingface.co/datasets/hltcoe/MultiVENT2.0)

audios_2025.tar.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:90a5faa0c0580c50c03df5ea6e3a67e739be34eb0e4644bdca9a82d902051dcb
+size 38665183

metadata_2025.tar.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6a38daa99329606877a6cef700445e69875dd78e89839820c2a7e4f93c65b811
+size 544692

videos_2025.tar.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f73fa64585658e025d4c2d35113a4a183553300d7d235082a1ef9dd96b7baf21
+size 656591186